Glossary

Short definitions to align terminology across docs and demos.

Agent (Compute Agent)

Lightweight program that executes bounded, deterministic operators. See: Compute Agent

Coordinates jobs and tasks across agents; validates results and handles retries. See: Orchestrator

Large unit of work decomposed into tasks by the orchestrator.

Smallest schedulable unit; deterministic, idempotent, bounded in time/memory.

Versioned set of deterministic primitives agents can execute. See: Operator Set v1

Registered/authenticated agent with reduced validation overhead; trust is revocable. See: Trust & Validation

Evidence generated by the orchestrator describing how a result was checked (method, rounds, outcome). See: API Examples

Scheduling action that excludes an agent after validation failures until manual review or timeout.

Advertised properties of an agent (CPU, memory, operator sets, protocol version, optional trust signals). See: Protocol

Optional mechanism to prove environment properties (e.g., TEEs); treated as advisory in early versions.

Any external location that stores tensors/datasets addressed by URI (e.g., file, HTTP, S3). See: Data Sources

An optional registry/signing service that returns handles and short‑lived access links for data sources. See: Data Sources

A reference to data (URI/handle) passed instead of embedding bytes in a task/result. Used to avoid moving large tensors.

Converting a reference or view into a concrete in‑memory tensor, typically for performance or subsequent random access.

A logical transformation over data (e.g., slice, transpose) that does not copy the underlying bytes.